Feeding MCMC output from one model into another

[aliso...@gmail.com]

Hi Nimble Devs,

Thank you all so much for your helpful hints and suggestions!

I have a spatial model within an integrated data model that causes impossibly long run times when the included in the integrated model. The integrated model already takes a while to run despite lots of optimizations. I have results of that spatial model fit to data that are not "integrated", but estimated posterior distributions of all of the spatial effects given the data outside of the joint model. I was thinking of possible solutions:  Feed the mcmc iterations of already sampled values from the spatial model directly into the MCMC sampler of the integrated model.  So, I am curious if it is possible to feed MCMC samples from one model run into a separate MCMC model run? I realize this is definitely not necessarily a good idea statistically, but we are at a impasse and I'm just knocking around ideas.

Thanks!

--Alison

[Daniel Turek]

Alison, I'll keep my response brief, but perhaps others will have more elaborate responses.

Short answer: without taking a stance on whether this "is a good idea statistically", yes, things like this can certainly be made to work.  Perhaps most simply by using a loop which iteratively stores new values into a model (values coming from the estimated posterior distributions of spatial effects using data outside the joint model), then executes an MCMC on the (updated) model for niter = 1 MCMC sampling iteration, and then repeats this process.  Does this sound like what you meant?  There are other ways you could approach the problem as well, but that's what comes to mind.

Cheers,

Daniel

--
You received this message because you are subscribed to the Google Groups "nimble-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nimble-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nimble-users/53b18590-74ff-4c7e-8e14-cc9a81c73ef6n%40googlegroups.com.

[Alison Ketz]

Thanks Daniel (and Thomas with an off list email),

We've also considered using the posterior as a prior to generate the effects as well. Any idea what you think might execute fastest for ~1000 spatial effects?